The training that works follows the same shape as the deployment: a short common baseline for everyone with a seat, then role-specific work in the tool each person actually holds, then a build program for the teams with a workflow worth redesigning. At Micron, that sequence moved self-rated Claude Cowork proficiency from 4.8 to 7.4 on a 10-point scale in four days.
This post covers the three deployment patterns we see, what to train at each level and for whom, what changed after training at Micron and International Airlines Group, and how to measure Claude adoption without relying on token consumption.
What's insideThree patterns account for nearly every Claude deployment in our discovery data, in this order of frequency.
The most common pattern: ChatGPT, Gemini or Copilot is the company-wide assistant, and Claude goes to the people whose work justifies it. A life-sciences company has ChatGPT for all 500 employees and Claude for roughly 100 of them, and its power users prefer Claude. An investment firm runs Gemini for general use and Claude for a super-user group of about a third of staff who do financial modelling, contract comparison and database work. An online marketplace holds Claude on 100 engineering licences while evaluating a second assistant for everyone else. A property operator gives Claude to directors and above and lets everyone else choose between two other tools. The allocation rule that comes up most often is scope of work, not grade.
Less common, and usually a deliberate choice: a software company rolled Claude out to roughly 4,000 employees in May 2026 and moved its engineers to Claude Code; a financial services firm has about 1,000 employees on Claude with foundational training complete; a security software company took Claude in enterprise form for the whole organisation specifically to protect its IP. At International Airlines Group, Correlation One ran a Claude enablement programme for about 50 AI champions and early adopters in October 2026, with a second cohort scheduled for early 2027.
A 20-person pilot in April that has grown by about five licences a week since; summer pilots followed by a six-week rollout to everyone; a firm that bought Claude Enterprise in August and runs its own monthly use-case sessions. In this pattern the licence count is driven by requests, which is why the first training question is usually "a lot of people don't know we have Claude."
Two cross-cutting facts: a large share of Claude deployments sit alongside Microsoft Copilot, and executives are frequently trained separately and first — one company's leadership team had private Claude training before the company-wide rollout began.
Match the training to the deployment pattern: one level for everyone with a seat, a second for the people who will build with it, a separate track for engineers, and a different session altogether for executives.
What is different about Claude under an Enterprise plan (data boundaries, Projects, file and data upload, artifacts), how to prompt with your own files rather than from a blank box, and responsible use for your industry. The gap that comes up most often in discovery is not awareness of Claude but what to do once a file is uploaded: people have the licence and do not know what to ask. At PharmaEssentia, the foundations sessions are built around one field-sales workflow — load a territory file, ask where the gaps are — so every participant leaves having completed a real task in the tool rather than a demonstration of features.
Workflow scoping, then building and testing in Claude Cowork and Projects with a coach per team. Micron's AI Impact Lab ran in person over four days in October 2026: one day scoping workflows with the AI Opportunity Engine, two days building in pods with coaches, and a final day testing, setting human-oversight points and presenting to stakeholders. IAG's Claude programme for champions and early adopters runs as a multi-week cohort with the same build-and-present structure.
Companies whose engineers use Claude Code ask for a separate curriculum on modern development practice, not a common floor; it came up repeatedly in discovery calls this autumn. The foundations session does not serve this group and should not try to.
Executives are directing what agents should do, not building them. The sessions that land show the back end of an agent — instructions, parameters, what governing it looks like — and the trade-offs in what an agent can and cannot do, so that when someone two levels down says they are building one, the leadership team knows what that means. At LivaNova, the senior-leadership sessions run on Copilot with a Claude breakout so leaders see both estates they own. Correlation One's executive AI training is built on that principle.
Proficiency moves within days when the training is hands-on in the participant's own workflow, and the gain is comparable whichever assistant the company has deployed.
Micron, four-day AI Impact Lab, October 2026. Self-rated Claude Cowork proficiency rose from 4.8 to 7.4 on a 10-point scale. The share of participants rating themselves 7 or higher rose from 27% to 76%. Every one of the eight capabilities measured rose from pre-lab to post-lab, and 95% of participants said they could apply the lab's method to another workflow without a coach.
International Airlines Group, Claude enablement programme, October 2026. The first session for roughly 50 champions and early adopters scored an NPS of +53 with 88% attendance, and a second cohort is scheduled for early 2027. IAG's earlier agent-building pilot, run on Copilot Studio from January to March 2026, produced 11 working agents that were demonstrated to about 110 people including Management Committee members. Same company, two tools, same method.
The tool is not the variable. Across Correlation One's 2026 post-engagement surveys, prompt-writing confidence rose by comparable amounts on Claude, on Copilot and on other enterprise assistants. What did affect results was whether participants had working access to the tool during the build, which is why licence provisioning is settled before day one of any Level 2 program.
Consumption tells you Claude is being used; it does not tell you whether work changed. One buyer in our discovery data put it plainly: they can always look at Claude consumption, but they do not consider it a good measure of productive use. Four measures do the job, and all four can be run from the company's own admin console without giving a training provider access to the tenant.
What not to use: completion rates for self-paced content, and consumption totals without a seat-activation denominator.
Correlation One delivers Claude Enterprise training at three levels — a live foundations session for everyone with a seat, a multi-day or multi-week build program in Claude Cowork and Projects for teams redesigning a workflow, and a separate Claude Code track for engineers — plus executive sessions on how agents are built and governed. 2026 programs include Micron's AI Impact Lab and International Airlines Group's Claude enablement programme.
How long does it take to train employees on Claude?Ninety minutes to two hours moves a general population from occasional use to prompting with their own files. Building a working agent or redesigned workflow takes four days in an intensive in-person format or up to seven weeks as a cohort with weekly sessions and a coach per team.
Should we train employees on Claude if most of the company uses ChatGPT or Copilot?Yes, for the population that holds a Claude seat, because the measured gains depend on practising in the tool the person actually has. Teach the method once across tools and run the hands-on work in each person's own assistant; a large share of the Claude deployments in our 2026 discovery calls run Claude next to Copilot.
Do engineers need separate training for Claude Code?Yes. Claude Code training covers modern development practice and agentic workflows in the codebase, which a general foundations session does not touch. Companies with engineers on Claude Code consistently asked for it as a distinct track in 2026.
Is Claude Enterprise training different from Copilot training?The structure is the same and the measured results are comparable. The content differs where the tools differ: Projects, file-based work and Cowork on Claude; in-app and tenant-grounded work in Word, Excel, Outlook and Teams on Copilot.
How do we know whether Claude training worked?Four measures from the company's own admin console and surveys: seat activation, usage in the 14 days after training versus the 14 days before, pre/post proficiency by capability, and the count of agents or workflows shipped by status. Token consumption on its own is not one of them.
A free AI Maturity score benchmarks your workforce's current AI capability and shows what to train first.
Get your free AI Maturity score → Talk to our team →